GitHub - ash-ai-org/ash-ai: Deploy Claude agents as production APIs — with sessions, streaming, sandboxing, and persistence handled for you.

npm install -g @ash-ai/cli

# Set your Anthropic API key
export ANTHROPIC_API_KEY=sk-ant-...

# Start the server
ash start

# Define an agent — it's just a folder
mkdir my-agent
cat > my-agent/CLAUDE.md << 'EOF'
You are a helpful coding assistant. Be concise and accurate.
When asked to write code, include working examples.
EOF

# Deploy and use it
ash deploy ./my-agent --name my-agent
ash session create my-agent
ash session send <SESSION_ID> "Write a prime number checker in Python"

Your agent folder          What you get back
─────────────────          ─────────────────────────────────────
my-agent/                  POST /api/sessions         → create session
├── CLAUDE.md              POST /api/sessions/:id/messages → stream messages (SSE)
├── .claude/               GET  /api/sessions/:id     → session status
│   ├── settings.json      POST /api/sessions/:id/pause  → pause session
│   └── skills/            POST /api/sessions/:id/resume → resume session
└── .mcp.json              DELETE /api/sessions/:id   → end session

my-agent/
├── CLAUDE.md                          # System prompt (required)
├── .claude/
│   ├── settings.json                  # Permissions — which tools the agent can use
│   └── skills/
│       └── search-and-summarize/
│           └── SKILL.md               # Reusable skill definitions
└── .mcp.json                          # MCP server connections (web fetch, memory, etc.)

# my-agent/CLAUDE.md
You are a customer support agent for Acme Corp.
Be friendly, cite the docs, and escalate billing issues to humans.

research-agent/
├── CLAUDE.md                          # "You are a research assistant..."
├── .mcp.json                          # Connect to web fetch, memory servers
└── .claude/
    ├── settings.json                  # Allow: Bash, WebSearch, mcp__fetch, mcp__memory
    └── skills/
        ├── search-and-summarize/
        │   └── SKILL.md              # /search-and-summarize skill
        ├── analyze-code/
        │   └── SKILL.md              # /analyze-code skill
        └── write-memo/
            └── SKILL.md              # /write-memo skill

import { AshClient } from '@ash-ai/sdk';

const client = new AshClient({ serverUrl: 'http://localhost:4100' });

const session = await client.createSession('my-agent');

for await (const event of client.sendMessageStream(session.id, 'Hello!')) {
  if (event.type === 'message') {
    process.stdout.write(event.data);
  }
}

await client.endSession(session.id);

from ash_sdk import AshClient

client = AshClient(server_url="http://localhost:4100")
session = client.create_session("my-agent")

for event in client.send_message_stream(session.id, "Hello!"):
    if event.type == "message":
        print(event.data, end="")

client.end_session(session.id)

CLI / SDK / Browser
        │
        │ HTTP + SSE
        ▼
┌──────────────────┐
│   ash-server     │──── SQLite / Postgres
│   (Fastify)      │
└────────┬─────────┘
         │ in-process
         ▼
┌──────────────────┐
│   SandboxPool    │
└──┬───────────┬───┘
   │           │
   ▼           ▼
┌──────┐   ┌──────┐
│Bridge│   │Bridge│    ← isolated child processes
│  1   │   │  2   │    ← Unix socket, newline-delimited JSON
└──┬───┘   └──┬───┘
   ▼           ▼
Claude SDK  Claude SDK